Doctoral Dissertations University of Connecticut Graduate School
7-27-2016
Essays on the Interactions of Consumer Behavior
and Firm Strategy in Multi-Channel Environments
Bin LiFollow this and additional works at:https://opencommons.uconn.edu/dissertations Recommended Citation
Li, Bin, "Essays on the Interactions of Consumer Behavior and Firm Strategy in Multi-Channel Environments" (2016).Doctoral Dissertations. 1209.
Essays on the Interactions of Consumer Behavior and Firm
Strategy in Multi-Channel Environments
Bin Li, PhD
University of Connecticut, 2016
Abstract
With the increasing popularity of the online channel, both consumers and firms are engaging in more and more multi-channel activities. On the one hand, consumers can integrate information from searches on both online and offline channels, and then decide on the best channel to buy from. On the other hand, firms need to consider consumer behavior in different channels in their strategy design. As a result, cross-channel interactions between consumer behavior and firm strategy can be within the same channel or across different channels. While the within-channel interaction has been studied extensively in the previous literature, there is much less research on the cross-channel interaction. In my dissertation, I add to the
understanding of consumer behavior and firm strategy in the multi-channel environment by empirically analyzing their cross-channel interactions.
This dissertation consists of three separate but related essays. The first answers the question: How does consumer behavior affect optimal product portfolio strategies in online versus offline channels? I develop an empirical model to simultaneously identify the
cannibalization effect (within a brand) and the competition effect (between different brands) in different retail channels. I further examine how these effects are affected by consumer
Bin Li – University of Connecticut, 2016
firm’s offline strategy affect consumer online behavior? I use a natural experiment to examine how the awareness and convenience effects from opening new retail stores affect the online search. The final essay answers the question: How does online banking affect entry/exit of offline bank branches? I develop and estimate a dynamic entry/exit model examining the relationship between technological advances and market structure evolution. My counterfactual analysis shows that the asymmetric reduction in operating costs is the most significant factor driving recent changes in the U.S. banking industry, followed by increased entry costs and increased deposits for large banks due to greater online presence. My findings provide important implications for firms engaging in multi-channel activities.
i
Essays on the Interactions of Consumer Behavior and
Firm Strategy in Multi-Channel Environments
Bin Li
B.S., Tsinghua University, 1991
A Dissertation
Submitted in Partial Fulfillment of the Requirements for the Degree of
Doctor of Philosophy at the
University of Connecticut
ii Copyright by
Bin Li
iii
APPROVAL PAGE
Doctor of Philosophy Dissertation
Essays on the Interactions of Consumer Behavior and Firm Strategy
in Multi-Channel Environments
Presented by Bin Li, B.S. Major Advisor _____________________________________________________________ Honju Liu Associate Advisor __________________________________________________________ Joseph Pancras Associate Advisor __________________________________________________________ Ting Zhu Associate Advisor __________________________________________________________ Nicholas Lurie Associate Advisor __________________________________________________________ William T. Ross, Jr. University of Connecticut 2016iv
Acknowledgments
This dissertation could not have been completed without all kinds of advice, guidance, and support from so many people over the past years. I wish to offer my heartfelt thanks to them.
First, I will be forever grateful to my five co-advisors: Dr. Hongju Liu, Dr. Joseph Pancras, Dr. Ting Zhu, Dr. Nicholas Lurie, and Dr. Bill Ross. They have guided me through the doctoral program using their experience in academia, professional knowledge in various areas, and endless encouragement. As the chairman of my doctoral committee, Dr. Liu means a lot to me: a lifetime friend, a gentle guiding hand, a watchful observer, and a silent but forceful motivator. The door to Dr. Pancras is always open to me, and he is patient to discuss and chat on anything I am interested in. The support and encouragement provided by Dr. Zhu cannot be overstated. Her unhesitant willingness to advise a doctoral outside of her institution merely for the sake of scholarship is an example worthy of long admiration. Dr. Ross inspired my research with his wonderful sense of humor and his eternal optimism. He is on top of my mind whenever I have difficulties in the PhD program. I would offer special thanks to Dr. Lurie. My journey in Marketing would not have started without you! Thank you for your continuous encouragement and suggestions in so many years.
I offer my sincere appreciation to Dr. Robin Coulter, Dr. Girish Punj, Dr. Susan Spiggle, Dr. Narasimhan Srinivasan, Dr. Jane Gu, Dr. Kunter Gunasti, Dr. David Norton, and Dr. Hee Mok Park for providing support and valuable suggestions to my research. In UConn Marketing, I have lots of good memories that I will remember forever.
v
I would like to take this opportunity to express my very deep gratitude to Dr. Xinxin Li from OPIM Department for her inspirations and support to make this dissertation possible. It is always my pleasure to work with her.
I would especially like to thank my family. To my mom, you understanding and endless love made me who I am today. My wife, Junjuan, has been extremely supportive of me
throughout this entire process. Thank you for your countless sacrifices to our family and help me get to this point. To my children, Julia and Kelly, you are the pride of my life, and the ultimate motivation to me in finishing my degree.
vi
Table of Contents
Chapter 1 Introduction ... 1
Chapter 2 Consumer Preferences, Cannibalization and Competition: Evidence from the Personal Computer Industry ... 4
2.1 Introduction ... 5
2.2 Literature Review ... 9
2.3 Data Description ... 11
2.4 Empirical Model ... 12
2.4.1 Cannibalization and Competition Effects ... 12
2.4.2 Consumer Preferences ... 16
2.5 Estimation ... 20
2.6 Results ... 22
2.6.1 Cannibalization and Competition Effects ... 22
2.6.2 Consumer Preferences ... 25
2.6.3 Discussion ... 28
2.7 Robustness Checks ... 32
2.8 Conclusions ... 33
Chapter 3 Retail Store Entry and Online Consumer Search: The Role of Awareness versus Convenience Effect ... 36
3.1 Introduction ... 37
3.2 Theoretical Framework ... 39
3.2.1 The Convenience Effect ... 40
3.2.2 The Awareness Effect ... 41
3.2.3 The Moderating Role of Consumer-Specific Characteristics ... 42
3.3 Data ... 44
3.3.1 Search Behavior ... 44
3.3.2 Store Entry and Location ... 45
3.3.3 Online Channel Efficiency ... 46
3.4 Research Methodology ... 47
3.4.1 Natural Experiment Design ... 47
3.4.2 Marketing and Shopping Regions ... 48
3.4.3 Model Specification ... 50
3.4.4 Matching ... 50
vii
3.5.1 Analysis on the Within Session Behaviors ... 59
3.6 Conclusion and Implications ... 60
3.6.1 Managerial Implications... 61
3.6.2 Limitations and Future Research ... 62
Chapter 4 Technology and Market Structure: An Empirical Analysis of Entry and Exit in the Banking Industry ... 64
4.1 Introduction ... 64
4.2 Literature Review ... 68
4.3 The U.S. Banking Industry ... 70
4.4 Data ... 72
4.4.1 Sample Markets... 72
4.4.2 Banks and Branches... 74
4.4.3 Broadband Penetration... 76 4.4.4 Descriptive Analysis ... 77 4.5 Model ... 78 4.5.1 Deposit Demand ... 79 4.5.2 Cost Structure ... 81 4.6 Empirical Strategy ... 86 4.6.1 Policy Functions ... 86 4.6.2 Estimator ... 87 4.6.3 Computational Details ... 89 4.7 Results ... 91
4.7.1 Deposit Demand Parameters ... 91
4.7.2 Structural Parameters for Entry and Exit Decisions ... 94
4.8 Counterfactual Simulations ... 96
4.9 Conclusion ... 100
Chapter 5 Conclusions and Future Research ... 102
Bibliography ... 103
Appendix 109 Appendix A. Selection of Nest Structure (Chapter Three) ... 109
viii
List of Figures
Figure 1. Interactions between consumers and firms in the multi-channel environment ... 1
Figure 2：Nest Structure of Our Two-level Nested Logit Model ... 13
Figure 3: Model Free Evidence of Higher Cannibalization Online ... 29
Figure 4：Comparison between Online and Offline Channels: Percentage Increase in Sales as the Number of Products Increases ... 31
Figure 5. Spatial distribution of the sample markets ... 73
Figure 6. Changes in the number of branches ... 75
ix
List of Tables
Table 1: Descriptive Statistics ... 20
Table 2: Correlation Matrix ... 21
Table 3: Estimation Results ... 23
Table 4: Cannibalization Parameter for Each Sub-Nest from Model M1 ... 24
Table 5: Competition Parameter for Each Nest from Model M1 ... 25
Table 6: Explaining Power of Loyalty and Search Variables ... 27
Table 7: Robustness Checks ... 32
Table 8: Data Sources ... 44
Table 9: Control and Treatment Groups ... 48
Table 10: Comparison of Geo-Information Before and After Full Matching ... 53
Table 11: Descriptive Statistics ... 53
Table 12: Correlation Matrix ... 55
Table 13: Number of Visits the Online Store ... 55
Table 14: Number of Pages per Visit ... 59
Table 15. Summary statistics of demographic variables ... 73
Table 16. Changes in the number of branches for all the markets ... 74
Table 17. Changes in the number of branches for the sample markets ... 75
Table 18. Broad band penetration ... 76
Table 19. HHI of bank level deposits ... 77
Table 20. Regression analyses on the number of branches and broadband penetration ... 77
Table 21. Estimates for deposit demand ... 92
x
Table 23. Counterfactual simulations for entry/exit probabilities ... 97 Table 24. Counterfactual simulations for the number of branches ... 98
1
Chapter 1
Introduction
The emergence and popularity of the online channel change the way consumers search for information and make decisions in online and offline environments. This has important implications for firms’ multi-channel strategies and well as the evolution of markets. On the one hand, consumers can integrate information from searches on both online and offline channels, and then decide on the best channel to buy from. On the other hand, firms need to consider consumer behavior in different channels in their strategy design. As a result, the cross-channel interactions between consumer behavior and firm strategy can be within the same channel or across different channels as shown in Figure 1.
Figure 1. Interactions between consumers and firms in the multi-channel environment
While the within-channel interaction has been studied extensively in the previous literature, there is much less research on the cross-channel interaction. In my dissertation, I add to the understanding of consumer behavior and firm strategy in the multi-channel environment by empirically analyzing their cross-channel interactions from three perspectives.
In the first essay, I develop an empirical model to simultaneously identify the
cannibalization effect (within a brand) and the competition effect (between different brands) in different retail channels. I further examine how these effects are affected by consumer
2
competition than the offline channel. Moreover, heterogeneity in consumer search behavior and brand loyalty explain a significant fraction of the variation in both cannibalization and
competition between the two channels. My findings suggest that firms should offer fewer models in the online channel and when consumers are highly loyal.
In the second essay, I examine how opening a new retail store affects the cognitive costs of online search and the physical cost of offline search. I find that, for consumers with prior experience on the retailer’s website, opening a new store leads to a 68% increase in the number of visits to the retailer’s website by consumers who live in the broader marketing area of the new store, while it leads to a 49% decrease in the number of visits to the retailer’s website by
consumers who live in the nearby shopping region. More interestingly, the effect of store entry in the shopping region on decreasing website search is weaker when consumers are more efficient in using the online channel. My results are robust to corrections for endogeneity in the choice of new store locations using different matching methods.
In the third essay, I develop and estimate a dynamic entry/exit model examining the relationship between technological advances and market structure evolution to understand the relationships between these intriguing phenomena. Despite the rise of consumer online banking, there has been little reduction in the number of brick and mortar bank branches in the U.S. At the same time, large national banks have expanded their branch networks at the cost of small local banks. My findings suggest that the advent of online banking provides significant competitive advantages to large national banks over small local banks that lower large banks’ offline operating costs and increase consumer deposits; yet increase entry costs. A counterfactual analysis shows that the asymmetric reduction in operating costs is the most significant factor
3
driving recent changes in the U.S. banking industry, followed by increased entry costs and increased deposits for large banks due to greater online presence.
In sum, the results presented in this dissertation provide insight into the multi-channel business models. These studies also contribute to the literature on product portfolio management and local competition.
4
Chapter 2
Consumer Preferences, Cannibalization and
Competition: Evidence from the Personal Computer Industry
Understanding the degree of cannibalization and competition in online and offline markets is important to firms’ product line designs. However, few empirical studies have measured both effects simultaneously or have examined the factors that determine the extent of cannibalization and competition. In this study, we develop an empirical model to identify
cannibalization and competition effects simultaneously in different markets, and further examine the impacts of consumer preferences on these two effects in a single integrated framework. Using data from the U.S. personal computer (PC) industry, we find that the online market exhibits stronger cannibalization and competition than the offline market. Both effects are significantly influenced by consumers’ search behavior and brand loyalty. Specifically, more active consumer search not only intensifies inter-brand competition but also amplifies intra-brand
cannibalization. In addition, search has a higher impact on cannibalization than competition. Stronger consumer brand loyalty mitigates inter-brand competition, but its effect on intra-brand cannibalization varies for different consumer segments. In markets consisting of more high-end consumers, the intra-brand cannibalization increases with consumer brand loyalty, while, in contrast, in markets consisting of more low-end consumers, the intra-brand cannibalization decreases with consumer brand loyalty. The differences in consumer search and brand loyalty explain a significant fraction of the variations in both cannibalization and competition between different PC markets.
5
2.1
Introduction
Product proliferation is an important component of a firm’s competitive strategy and is commonly observed in practice. For example, in the personal computer (PC) industry, we often observe multiple models of desktops or laptops offered by the same firm or brand. In 2001, HP offered 19 desktop models with varying configurations, and this number increased to 35 in 2008. Similar phenomena were observed for other PC manufacturers as well. Product proliferation is not specific to the PC industry. According to Bernard et al. (2006), multiproduct firms account for more than 90% of the output in the U.S. manufacturing sectors.
Product proliferation has at least two competing effects on firm profitability.1 First, it has
an inter-brand competition effect: when introducing a greater variety of product offerings, a firm can attract new consumers with heterogeneous tastes and induce consumers to switch from competitors (Bayus and Putsis 1999). Second, it has an intra-brand cannibalization effect: products offered by the same firm are often considered by consumers as close substitutes so that “one product’s customers are at the expense of other products offered by the same firm” (Mason and Milne 1994, p.163).2
Firms need to consider both intra-brand cannibalization and inter-brand competition effects when designing their product lines (Kadiyali et al. 1998; Ruebeck 2005; Wilson 2011). A stronger cannibalization effect gives incentives for firms to shorten their product lines, whereas a fiercer competition between firms may induce them to expand their product lines. Due to the rapid growth of the online market, the product line design in online and offline markets becomes
1 Other effects such as market expansion are not the focus of this study, but will be controlled in our
analysis.
2 In this study, we assume each firm has a single brand, so we use brand and firm interchangeably. Within
6
an increasingly more important issue for multiproduct firms to consider. Earlier work in information systems suggests some important differences between the online and offline markets. For example, the online channel has lower search cost (Bakos 1997; Clemons et al. 2002; Ghose and Yao 2011) and menu cost (Bailey 1998; Brynjolfsson and Smith 2000) compared to brick-and-mortar stores. These differences have important implications to firms’ online pricing (Clemons et al. 2002; Ghose and Yao 2011). Recent studies have also started to examine the implications of this new channel to firms’ product strategy. For example,
Brynjolffson et al. (2011) suggest that the online market creates a long tail in the distribution of sales by substantially increasing the market shares of niche products. A natural question to ask following these findings is: how should firms optimize their product offerings online? Should firms offer more products online than offline given its ability to attract consumers into buying more niche products? Given that the online and offline markets differ in many key characteristics that can potentially affect the degree of intra-brand cannibalization and inter-brand competition, it is important for firms to gauge both cannibalization and competition effects in online and offline markets and to understand their influencing factors to make optimal product line decisions.
Our first objective in this study is to empirically measure intra-brand cannibalization and inter-brand competition and examine whether they vary across online and offline markets. We develop a unified framework to jointly estimate both cannibalization and competition in different markets in the U.S. PC industry. We find significant differences in both effects between online and offline markets. Specifically, the online market exhibits both stronger intra-brand
7
This leads to our second research question: what factors drive intra-brand
cannibalization and inter-brand competition in different markets? In particular, we focus on how consumer brand loyalty and search behavior influence cannibalization and competition,
respectively. To answer this question, we model both intra-brand cannibalization and inter-brand competition as functions of consumer preferences. We find that, consistent with intuition,
stronger consumer brand loyalty reduces inter-brand competition. The impact of brand loyalty on intra-brand cannibalization, however, varies for different segments of consumers. Specifically, stronger brand loyalty increases intra-brand cannibalization in markets consisting of more high-end consumers (high income, low price sensitivity), but decreases intra-brand cannibalization in markets consisting of more low-end consumers (low income, high price sensitivity). In addition, we find that more active consumer search not only intensifies inter-brand competition but also amplifies intra-brand cannibalization to a larger extent, suggesting that more consumer search before purchase encourages more comparisons within the same brand than between different brands. The differences in consumer search and brand loyalty explain a significant fraction of the variations in both cannibalization and competition between different PC markets.
This study makes the following important contributions. First, our study offers a
framework to jointly identify intra-brand cannibalization and inter-brand competition in different markets as well as the impacts of demand side factors on competition and cannibalization in a single integrated model, while at the same time controlling for potential endogeneity issue. Prior studies (e.g., Brynjolfsson et al. 2009; Hui 2004; Watson 2009) have focused on either
cannibalization or competition, but have not measured them simultaneously for different markets. Our results suggest that both cannibalization and competition vary across online and offline markets and need to be considered jointly for optimal product line design. Specifically,
8
while the higher online competition between firms has been recognized in academic research and practice, the higher online cannibalization between products within a firm has been largely overlooked. This could lead to sub-optimal product line design in the online market.
Second, our study highlights important demand side factors (i.e., consumer preferences) that influence both cannibalization and competition, whereas prior studies (e.g., Thomadsen 2007) have mainly focused on the supply side factors (e.g., product characteristics and pricing). Our results suggest that consumer brand loyalty and search behavior have significant impacts on cannibalization and competition, and in addition, that the differences in these consumer
preferences play an important role in explaining the variations in cannibalization and competition between different markets.
In particular, when drawing implications about the impact of brand loyalty on product cannibalization, it is important for firms to consider consumers’ brand loyalties in different segments because their impacts are different. For example, if a market is mainly composed of loyal consumers from the high-end segment, intra-brand cannibalization can be high and it may be optimal for manufacturers to offer fewer products to the market. In fact, in our data, we observe a higher interaction between loyalty and income in the online channel. Because brand loyalty of high end consumers has a positive impact on cannibalization, the characteristics of the online population in the PC market can make it more likely for us to observe a higher
cannibalization online.
While prior IS studies have primarily focused on the impact of lowered search cost on inter-firm competition, we find that the lowered search cost not only affects competition, but also affects cannibalization. More importantly, our result suggests that search in fact can have a higher impact on cannibalization than competition. This result has important implications: while
9
the increased competition due to lower search cost online may encourage firms to offer more products to attract consumers, it can increase intra-brand cannibalization to a larger extent, limiting the optimal number of products to offer. Therefore, the lowered search cost online does not necessarily favor a longer product line. This can be counter-intuitive since online stores are often not limited by shelf space or capacity constraints, and it is generally cheaper to host a greater product variety online. However, a shorter product line online may be optimal if the higher intra-brand cannibalization online is a major concern.
2.2
Literature Review
Our study is closely related to the literature on product proliferation. Most studies in this research stream develop theoretical models to analyze firms’ decisions on product quality (Desai 2001; Katz 1984; Moorthy 1984) or on the length of product line (Bayus and Putsis 1999; Bordley 2003; Liu and Cui 2010; Shugan 1989). When consumers have heterogeneous
preferences over quality, firms have incentives to vertically differentiate their products to meet the needs of different consumers. Desai (2001) develops a model for duopolistic competition where consumers have heterogeneous preferences over brand in addition to quality. His results suggest that the brand preferences of consumers with different quality sensitivities affect intra-brand cannibalization differently. Studies on the length of product line suggest that a longer product line can lead to more severe product cannibalization and diminishing marginal contribution to firm performance (Draganska and Jain 2005) or even negative impact on firm profit (Bayus and Putsis 1999). The effectiveness of product proliferation as a marketing strategy thus is clearly limited by the extent of product cannibalization. However, there have been few attempts to study cannibalization, competition and their influencing factors simultaneously, which is a gap in the literature that this study aims to fill.
10
Prior studies have, however, studied each of the two effects (cannibalization and competition) separately. The research stream on product cannibalization has used different methods to quantify the extent of intra-brand cannibalization. Most papers infer about cannibalization indirectly from cross-product price elasticity (Berry et al. 1995), or from the change of firm performance in response to the change in the length of product line (Bayus and Putsis 1999; Draganska and Jain 2005). Notably Hui (2004) examines a direct measure of intra-brand cannibalization and then links cannibalization to firm intra-brand value. Compared with the cross elasticity between individual products, this measure indicates the degree of cannibalization at an aggregate level. Such a single measure is especially suitable to study our research question, which examines the link between cannibalization and consumer characteristics at the market level. We thus use similar aggregate-level measures in this study. Our study is different from Hui (2004) in two aspects: first, we measure not only intra-brand cannibalization but also inter-brand competition, and second, more importantly, we further investigate the impact of consumer preferences on both cannibalization and competition in a single integrated model.
The recent research stream on inter-brand competition effect focuses on comparing this effect between online and offline markets. Most research in this area focuses on the
consequences of lower search costs online and often uses price dispersion as an index of competition (e.g., Baye et al. 2006; Brynjolfsson et al. 2010; Clemons et al. 2002). Some researchers suggest that firms should optimize their product line design by offering more
differentiated products in the online market to mitigate online competition (Bar-Isaac et al. 2009; Kuksov 2004). We capture the effect of consumer search on product competition in our model, and different from previous studies, we further compare this effect with the impact of consumer
11
search on cannibalization and include consumer preference variables to help explain the differences in both cannibalization and competition between different markets.
2.3
Data Description
Our empirical analyses focus on the U.S. PC industry. We choose the PC industry as the empirical context for the following three reasons. First, it is a highly competitive market with a number of major PC makers, each offering a variety of computer models. This combination allows us to study both inter-brand competition and intra-brand cannibalization. Second, products in the PC industry are vertically differentiated with similar functions, so PC vendors compete fiercely on product quality, such as CPU speed. This enables us to apply Lancaster’s characteristics theory of value (Lancaster 1966) in our empirical model. Third, this is a dynamic industry characterized by frequent new product introductions and quality improvements, which provide us with more variations in the data that are helpful for parameter identification.
Our data come from two datasets. The first dataset is PC sales data between 2004 and 2008 collected by the International Data Corporation (IDC). For each PC model, IDC records its manufacturer, form factor (desktop or laptop), CPU, average retail price, and unit sales in each quarter. The sales data are available for six distribution channels: Internet, retail, direct inbound, direct outbound, dealer/value-added reseller/systems integrator, and others. In this study, we use data for the Internet and retail channels only,3 because of two reasons: first, our focus is on
comparing the two channels, and second, these two channels mostly reflect consumer purchases and therefore can be matched with the information from our second dataset—the Consumer Technographics Benchmark survey data for the same period from Forrester Research. The
3 Consumer purchases in other channels will fall under an outside option in our demand model presented
12
benchmark survey is an annual survey of over 50,000 households in North America, from which the households from Canada are excluded from our analysis. The survey includes questions about consumer preferences, general shopping behaviors and purchase decisions for computers, among other products. Detailed information about consumers’ PC purchases in the survey is used to match their shopping behaviors to IDC sales data. Specifically, our matching is based on three factors: year of purchase, form factor (desktop or laptop) and channel of purchase (online or offline), given that our consumer preference variables will be measured at the market level. While the IDC sales data allow us to estimate a demand model that quantifies cannibalization and competition in the PC market, the Forrester survey data help explain how consumer preferences affect the extent of cannibalization and competition in the same market.
In our study, we restrict our analyses to firms identified in both the IDC sales data and the Forrester surveys. Our final dataset includes the eight largest PC firms between 2004 and 2008: Acer, Apple, Dell, Gateway, HP, Lenovo, Sony, and Toshiba. These firms accounted for 77% of total sales in the PC industry during the period. In addition, we aggregate the quarterly IDC sales data to a yearly level to match the annual Forrester surveys.
2.4
Empirical Model
2.4.1
Cannibalization and Competition Effects
To address our first research objective that aims to identify both intra-brand
cannibalization and inter-brand competition and compare them across online and offline markets, we develop an empirical model under the generalized extreme-value (GEV) discrete choice framework (McFadden 1978). In particular, we consider consumers’ PC purchases for different form factors (desktop or laptop) and in different channels (online or offline). We thus set up a
13
two-level nested logit model as illustrated in Figure 24. In the first level, the four nests
correspond to the four markets identified by form factor and channel combinations: laptop-online, desktop-laptop-online, laptop-offline, and desktop-offline. The second level sub-nests include all the firms competing in the same market, and each sub-nest has all the computer models from the same firm. With eight firms in our analysis, in total we have 30 sub-nests in the second level because Toshiba desktops are excluded due to very limited sales.
Figure 2：Nest Structure of Our Two-level Nested Logit Model
Let l index form factor and m index channel. The utility to consumer h from purchasing PC model i from firm j in nest lm is specified as:5
𝑈ℎ𝑖𝑗𝑙𝑚 = 𝛼𝑝𝑝𝑖𝑗𝑙𝑚+ 𝜑1𝑎𝑖𝑗𝑙+ 𝜑2𝑎𝑖𝑗𝑙2 + 𝑏𝑗+ 𝑐𝑖+ 𝑑𝑙𝑚+ 𝜉𝑖𝑗𝑙𝑚+ 𝜀ℎ𝑖𝑗𝑙𝑚 (1) Here, 𝑝𝑖𝑗𝑙𝑚 is the price of the PC model.𝑎𝑖𝑗𝑙represents the age of the PC model, measured as the time elapsed since it was first available in either channel. We include the quadratic term of 𝑎𝑖𝑗𝑙 to capture any nonlinearity in consumer preferences for product vintage. 𝑏𝑗
is the brand value of firm j. 𝑐𝑖 captures consumer preferences for the CPU used in model i, which
4 The rationale behind the choice of our nest structure is explained in detail in Appendix A. 5 For readability purpose, we suppress all time subscripts.
14
is a key indicator for PC quality because it is the most important component in determining the performance of a PC (Bresnahan et al. 1997; Hui 2004; Venkatesh and Mahajan 1997).6𝑑𝑙𝑚
stands for the interactions of the form factor dummies with the channel dummies to capture form factor and channel fixed effects. 𝜉𝑖𝑗𝑙𝑚 captures the unobserved product attributes, while 𝜀ℎ𝑖𝑗𝑙𝑚
captures the idiosyncratic taste of consumer h.
The unconditional market share of PC model i from firm j in nest lm is the product of three terms:
𝑠𝑖𝑗𝑙𝑚 = 𝑠𝑖|𝑗|𝑙𝑚∙ 𝑠𝑗|𝑙𝑚∙ 𝑠𝑙𝑚. (2)
Here 𝑠𝑙𝑚 is the (unconditional) market share of nest lm, 𝑠𝑗|𝑙𝑚 is the market share of firm j
in nest lm, and 𝑠𝑖|𝑗|𝑙𝑚 is the market share of model i within firm j in nest lm. Following the generalized extreme-value framework, we have:
𝑠𝑖|𝑗|𝑙𝑚 =𝑒𝛿𝑖𝑗𝑙𝑚 1−𝜎𝑗𝑙𝑚 ⁄ 𝐺𝑗𝑙𝑚 ; (3) 𝑠𝑗|𝑙𝑚 = 𝐺𝑗𝑙𝑚1−𝜎𝑗𝑙𝑚 1−𝜌𝑙𝑚 ⁄ 𝐹𝑙𝑚 ; (4) 𝑠𝑙𝑚 = 𝐹𝑙𝑚1−𝜌𝑙𝑚 ∑𝑙𝑚𝐹𝑙𝑚1−𝜌𝑙𝑚; (5) where 𝛿𝑖𝑗𝑚 = 𝛼𝑝𝑝𝑖𝑗𝑙𝑚+ 𝜑1𝑎𝑖𝑗𝑙+ 𝜑2𝑎𝑖𝑗𝑙2 + 𝑏𝑗 + 𝑐𝑖 + 𝑑𝑙𝑚 + 𝜉𝑖𝑗𝑙𝑚; (6) 𝐺𝑗𝑙𝑚 = ∑ 𝑒𝛿𝑖𝑗𝑙𝑚⁄1−𝜎𝑗𝑙𝑚 𝑖∈𝑗𝑙𝑚 ; (7) 𝐹𝑙𝑚 = ∑ 𝐺𝑗𝑙𝑚1−𝜎𝑗𝑙𝑚⁄1−𝜌𝑙𝑚 𝑗∈𝑙𝑚 . (8)
As a result, the unconditional market share of model i from firm j in nest lm is given by:
6 Because there are over 100 different CPUs in our dataset, it is empirically challenging to estimate a
separate dummy variable for each CPU. Following Gordon (2009) and Sriram et al. (2010), we classify all CPUs into five categories based on their benchmark scores. Our results are robust to the number of CPU categories.
15 𝑠𝑖𝑗𝑙𝑚 = 𝑒𝛿𝑖𝑗𝑙𝑚 1−𝜎𝑗𝑙𝑚 ⁄ 𝐺𝑗𝑙𝑚1−(1−𝜎𝑗𝑙𝑚) (1−𝜌𝑙𝑚) ⁄ ∙𝐹𝑙𝑚𝜌𝑙𝑚∙∑𝑙𝑚𝐹𝑙𝑚1−𝜌𝑙𝑚 . (9)
Note that 𝐺𝑗𝑙𝑚 and 𝐹𝑙𝑚 are the inclusive values that consumers expect to receive when choosing the best option in sub-nest jlm and in nest lm,respectively. They indicate the overall attractiveness of the corresponding sub-nest or nest.
Following Berry (1994), we can transform Equation (9) into its equivalent linear form:
log(𝑦𝑖𝑗𝑙𝑚) − log(𝑦0) = 𝛼𝑝𝑝𝑖𝑗𝑙𝑚+ 𝜑1𝑎𝑖𝑗𝑙+ 𝜑2𝑎𝑖𝑗𝑙2 + 𝑏𝑗 + 𝑐𝑖+ 𝑑𝑙𝑚 +
𝜌𝑙𝑚log(𝑆𝑗|𝑙𝑚) + 𝜎𝑗𝑙𝑚log(𝑆i|𝑗|𝑙𝑚) + 𝜉𝑖𝑗𝑙𝑚. (10)
Here, 𝑦𝑖𝑗𝑙𝑚 is the unit sales of model i, and 𝑦0 is the quantity corresponding to the outside option. To allow for the possibility that the market size may be changing over time, we use a set of yearly dummies to control for the size of outside option in our estimation process.7
The parameter 𝜎𝑗𝑙𝑚 measures the similarity between the alternatives within sub-nest jlm
perceived by consumers, and thus captures the intra-brand cannibalization within firm j in market
lm. The rationale is that when consumers choose products to maximize their utility, the extent of competition between products depends on how closely these products are perceived by
consumers. The larger the similarity parameter 𝜎𝑗𝑙𝑚, the closer the products are perceived, and the stronger the cannibalization between firm j’s own PC models. This can be seen from
Equation (10) above. Note that log(𝑆i|𝑗|𝑙𝑚) is always negative in the 𝜎𝑗𝑙𝑚log(𝑆i|𝑗|𝑙𝑚) term. When 𝜎𝑗𝑙𝑚 becomes larger, the sales of each individual product (𝑦𝑖𝑗𝑙𝑚) will become smaller given the same market shares within sub-nest jlm, indicating stronger intra-brand
7 Using our regression results, we also examine the evolution of the outside option in our model. We
compare the estimated yearly dummies with the total number of PC models over the years. We find that as the number of models increased, the size of outside option decreased over the years. This pattern may provide a crude indication for the market expansion effect. However, given the focus of this study, we do not pursue to investigate this effect in detail.
16
cannibalization. Following the same logic, we consider the parameter 𝜌𝑙𝑚 to capture the inter-brand competition between firms in market lm since it measures the perceived similarity between the alternatives within nest lm. The larger the parameter 𝜌𝑙𝑚, the more intense the competition between different firms within the market.8
To compare cannibalization and competition across online and offline markets, we model both 𝜎𝑗𝑙𝑚 and 𝜌𝑙𝑚 as functions of the interacted formfactor and channel dummies. We also include firm fixed effects in the function for 𝜎𝑗𝑙𝑚. To ensure that 𝜎𝑗𝑙𝑚 and 𝜌𝑙𝑚 are between 0 and 1, we apply a logistic transformation:
𝜎𝑗𝑙𝑚 =
𝑒𝜃𝑗+𝜃𝑙𝑚
1+𝑒𝜃𝑗+𝜃𝑙𝑚; (11)
𝜌𝑙𝑚 = 𝑒𝜇𝑙𝑚
1+𝑒𝜇𝑙𝑚. (12)
After estimating 𝜎𝑗𝑙𝑚 and 𝜌𝑙𝑚, we can then compare these two effects between online and offline markets to examine whether the cannibalization and competition effects vary in different markets.
2.4.2
Consumer Preferences
To answer our second research question that aims to examine the impacts of consumer preferences on cannibalization and competition, we further link 𝜎𝑗𝑙𝑚 and 𝜌𝑙𝑚 to variables indicating consumer preferences. Literature suggests that there are at least two aspects of
8 To see the connection between our cannibalization parameter and change in sales, we run a simple
simulation according to Equation (10) to see how product sales respond to changes in the number of products at different values of 𝜎𝑗𝑙𝑚. By fixing other parameters except the number of products offered by a firm, it shows that,
increasing the number of products, e.g., from 5 to 10, doubles the total sales for the firm when 𝜎𝑗𝑙𝑚 = 0, but leads to
no gain in total sales for the firm when 𝜎𝑗𝑙𝑚= 1. Similar insight can be drawn for competition parameter 𝜌𝑙𝑚. The
17
consumer preferences related to intra-brand cannibalization and inter-brand competition: brand loyalty and search behavior.
Multiproduct firms usually differentiate products on quality to serve consumer segments varying in preferences for quality (Moorthy 1984; Moorthy 1988; Mussa and Rosen 2012). Firms are concerned, however, about their low-quality product cannibalizing the demand meant for their high-quality product. The magnitude of this cannibalization is determined by the
substitutability between high-quality and low-quality products. This substitutability is contingent on the price and quality of each product, while firms’ price and quality decisions are affected by consumers’ preferences for different brands. The brand loyalties of price-sensitive (low-end) and quality-conscious (high-end) consumers can have different effects on firms’ price and quality decisions, which, as a result, affect intra-brand cannibalization differently (Desai 2001). Based on the results from Desai’s analytical model, if the brand preference of price-sensitive consumers is higher, the lack of competition between low-quality products from different firms will likely make them less attractive for the high-end consumers (as a result of firms optimizing both quality and price), thus mitigating intra-brand cannibalization. Conversely, if the brand preference of quality-conscious consumers is higher, the lack of competition between high-quality products will likely make them less attractive and thus increase the high-end consumers’ incentives to buy the low-quality product, intensifying the intra-brand cannibalization.
Correspondingly, we expect the impact of brand loyalty on intra-brand cannibalization to be positive in markets that consist of more quality-conscious consumers and negative in markets that consist of more price-sensitive consumers.
In addition to affecting intra-brand cannibalization, consumer brand loyalty also affects the competition between brands. Consumers loyal to a brand typically conduct fewer searches for
18
alternatives, effectively reducing their consideration set and leading to less brand switching (Sambandam and Lord 1995). Moreover, loyal customers become less price sensitive
(Krishnamurth and Raj 1991) and less responsive to promotions (Empen et al. 2011; Raju et al. 1990). As a result, stronger brand loyalty can reduce the intensity of inter-brand competition, which applies to both high-end and low-end consumer segments. Due to a lack of theoretical support and empirical evidence, we do not consider the difference between high-end and low-end consumers when modeling the effect of brand loyalty on inter-brand competition.
Consumer search is another factor that may affect both cannibalization and competition. Consumers usually engage in active prior-purchase search when they are uncertain about the attributes or prices of alternative products and must gather further information in order to reach a utility-maximizing choice (Feinberg and Huber 1996; Ratchford 1982). The prior-purchase search is more important for complex decisions, such as purchases of computers (Bettman 1979). After filtering the available alternatives, consumers form their consideration sets that include products that they will consider (Wright and Barbour 1977). Because prior-purchase search may involve comparing products from different firms as well as comparing products from the same firm, it has the potential to affect inter-brand competition as well as intra-brand cannibalization.
To capture the impacts of brand loyalty and search behavior on cannibalization and competition, we further incorporate these consumer preference variables into the functions for
𝜎𝑗𝑙𝑚and 𝜌𝑙𝑚: 𝜎𝑗𝑙𝑚 = 𝑒𝛽1𝐿𝑙𝑚+𝛽2𝐼𝑙𝑚+𝛽3𝐿𝑙𝑚∗𝐼𝑙𝑚+𝛽4𝐿𝑙𝑚∗𝐼𝑙𝑚2 +𝛽5𝑆𝑙𝑚+𝛽6𝑀𝑗𝑙𝑚+𝜃𝑗+𝜃𝑙𝑚 1+𝑒𝛽1𝐿𝑙𝑚+𝛽2𝐼𝑙𝑚+𝛽3𝐿𝑙𝑚∗𝐼𝑙𝑚+𝛽4𝐿𝑙𝑚∗𝐼𝑙𝑚2 +𝛽5𝑆𝑙𝑚+𝛽6𝑀𝑗𝑙𝑚+𝜃𝑗+𝜃𝑙𝑚 ; (13) 𝜌𝑙𝑚 = 𝑒𝛾1𝐿𝑙𝑚+𝛾2𝑆𝑙𝑚+𝜇𝑙𝑚 1+𝑒𝛾1𝐿𝑙𝑚+𝛾2𝑆𝑙𝑚+𝜇𝑙𝑚. (14)
19
Here, 𝐿𝑙𝑚 is the average brand loyalty of consumers who shop for form factor l in channel m. We use the item “When I find a brand I like, I stick to it” in the Forrester survey to
measure brand loyalty.9 In the literature, brand loyalty has been operationalized from either
behavioral or attitudinal perspectives. Specifically, behavioral loyalty is determined by the observed outcome, such as repeated purchase (Guadagni and Little 1983; Kahn et al. 1986), while attitudinal loyalty focuses on consumers’ stated preferences and purchase intentions (Chaudhuri and Holbrook 2001; Jacoby and Kyner 1973). Our measurement falls into the latter category.
𝐼𝑙𝑚 represents the average income of consumers who shop for form factor l in channel m. Consumers’ income levels reflect their price sensitivities. High-income (low-income) consumers are expected to be low (high) in price sensitivities. Thus, if a market has a high (low) average income, we expect it to be mainly influenced by high-end (low-end) consumers. Given our earlier discussion that the impact of brand loyalty on intra-brand cannibalization differs across markets consisting of different consumers, we interact the brand loyalty 𝐿𝑙𝑚 with the average income 𝐼𝑙𝑚 in both linear and quadratic terms, so that the impact of brand loyalty on
cannibalization may vary at different income levels.
𝑆𝑙𝑚 captures the average consumer preference for prior-purchase search for the consumers who shop for form factor l in channel m. A consumer’s preference for search is
measured by the survey item about consumer tendency to “research products for purchase” in
the Forrester survey. To control for the impact that the number of models has on product
9 We also tried using a different survey item “I would pay more for products consistent with an image I
20
cannibalization, we add 𝑀𝑗𝑙𝑚 into our model, which is the ratio of the number of models offered by firm j to the total number of models in nest lm.
2.5
Estimation
We first replace 𝜎𝑗𝑙𝑚 and 𝜌l𝑚 in Equation (10) with Equations (11) and (12) to answer our first research question of whether cannibalization and competition vary across markets, and then replace 𝜎𝑗𝑙𝑚 and 𝜌𝑙𝑚 in Equation (10) with Equations (13) and (14) to answer our second research question of how consumer preferences affect cannibalization and competition. This setup jointly estimates cannibalization and competition as well as the impacts of two demand side factors (search and loyalty) on both cannibalization and competition in a single integrated framework. This approach is more efficient than a two-step procedure like the one used by Hui (2004). This setup also allows us to observe the extent of variations in cannibalization and competition that can be explained by the consumer preference variables included in this study. Table 1 presents the summary statistics of the key variables and Table 2 presents their
correlations.
Table 1: Descriptive Statistics
Description Obs. Mean St. Dev. Min Max
𝑦𝑖𝑗𝑙𝑚 Unit sales 2,851 65,691 154,819 2 2,160,372
𝑠𝑗|𝑙𝑚 Market share of firm j in nest 𝑙𝑚 2,851 0.20 0.24 0.00* 0.89
𝑠𝑖|𝑗|𝑙𝑚 Market share of model i in sub-nest 𝑗𝑙𝑚 2,851 0.05 0.08 0.00* 0.98
𝑝𝑖𝑗𝑙𝑚 Price (in $1,000) 2,851 0.92 0.48 0.21 5.00
𝑎𝑖𝑗𝑙 Product age (in years) 2,851 0.60 1.01 0.00 11.23
𝐿𝑙𝑚 Consumer loyalty 2,851 3.52 0.06 3.36 3.62
𝑆𝑙𝑚 Consumer search before purchase 2,851 0.16 0.04 0.12 0.29
𝑀𝑗𝑙𝑚 Percent of models offered by firm j 2,851 0.21 0.10 0.02 0.56
𝐼𝑙𝑚 Consumer income (in $10,000) 2,851 5.84 1.04 4.48 8.19
21
Table 2: Correlation Matrix
𝑦𝑖𝑗𝑚𝑡 𝑆𝑖|𝑙𝑚 𝑆𝑖|𝑗|𝑙𝑚 𝑝𝑖𝑗𝑙𝑚 𝑎𝑖𝑗𝑙 𝐿𝑙𝑚 𝑆𝑙𝑚 𝑀𝑗𝑙𝑚 𝐼𝑙𝑚 𝑦𝑖𝑗𝑙𝑚 1.0000 𝑆𝑗|𝑙𝑚 0.3885 1.0000 𝑆𝑖|𝑗|𝑙𝑚 0.3426 -0.1226 1.0000 𝑝𝑖𝑗𝑙𝑚 -0.0591 -0.1155 0.1068 1.0000 𝑎𝑖𝑗𝑙 0.0725 0.0404 0.1763 -0.0559 1.0000 𝐿𝑙𝑚 0.0628 0.0911 -0.0473 -0.0699 -0.0402 1.0000 𝑆𝑙𝑚 -0.0887 0.0108 0.0976 0.0911 0.0745 -0.2169 1.0000 𝑀𝑗𝑙𝑚 -0.0026 0.3239 -0.3376 -0.2836 -0.0023 0.1711 0.0131 1.0000 𝐼𝑙𝑚 0.0514 0.0740 0.2056 0.1517 0.1371 -0.2114 0.6399 0.1763 1.000
Note that in Equation (10), the price 𝑝𝑖𝑗𝑙𝑚 may correlate with the error 𝜉𝑖𝑗𝑙𝑚, because firms’ pricing decisions can be based on some information in 𝜉𝑖𝑗𝑙𝑚 that is unobserved to the econometrician. Moreover, market shares may also be affected by the unobserved product
attributes in 𝜉𝑖𝑗𝑙𝑚. As a result, both price and market shares can be endogenous in our model. We use two sets of instruments to ensure consistent estimates of model parameters. Our first set of instruments uses supply-side cost shifters. Standard differentiated-product models predict that price is a function of marginal cost, so input prices are often used as instrumental variables for prices of end products by prior studies (e.g., Draganska and Jain 2005; Chu et al. 2007).
Specifically, we use the Producer Price Index (PPI) for CPUs as an instrument. Given that CPU is a key input to PC production, the PPI of CPU should be highly correlated with PC prices. On the other hand, consumer demand for PCs should not be directly affected by CPU prices after accounting for PC prices. As a result, the PPI of CPU should not be correlated with the error term in the demand system for PC, making it a valid instrument. We download the PPI data for the product category of “Microprocessors” from BLS’s website. Our second set of instruments is derived from the observed attributes of related PC models. Berry (1994) shows that the observed
22
attributes of related products are valid instruments for price and within-nest market share in discrete choice demand models for differentiated products because these attributes are
predetermined, or at least determined before consumers’ evaluations of the unobserved product attributes are revealed. This type of instruments has been widely used by prior studies (e.g., Berry 1994; Berry et al. 1995; Hui 2004; Zhu and Zhang 2010). Specifially, we use: (1) the sum of product age for the other PC models from the same firm with the same form factor in the same channel; (2) the sum of product age for the PC models from all the other firms with the same form factor in the same channel. To ensure that we have a sufficient number of instruments, we interact these two sets of instruments with the combination of firm, form factor, and channel dummies during the estimation process (Chu et al. 2007; Hui 2004). We have multiple
endogenous variables and therefore multiple first-stage regressions. The average first-stage R2 is
0.6942 and adjusted-R2 is 0.6813.
A linear model with instrumental variables is typically estimated using a standard two-stage least squares (2SLS) procedure. However, given the logistics transformations applied on
𝜎𝑗𝑙𝑚 and 𝜌𝑙𝑚, the 2SLS estimation procedure is not applicable in our nonlinear setup. Therefore, we use a generalized method of moments (GMM) estimator. The detail of the GMM estimator is in Appendix B.
2.6
Results
2.6.1
Cannibalization and Competition Effects
Table 3 shows our estimation results. Model M1 corresponds to the model specified in Equation (10) with 𝜎𝑗𝑙𝑚 and 𝜌𝑙𝑚 replaced with Equations (11) and (12) respectively. The
23
parameter estimates allow us to calculate the average intra-brand cannibalization and the average inter-brand competition for both online and offline markets.
Table 3: Estimation Results
M1 M2 M3 Description No Consumer Preference Variables (GMM) Main Model (GMM)
Linear Link Function (2SLS)
Utility function
𝛼𝑝(𝑝𝑖𝑗𝑙𝑚) Price -0.8291 *** -0.3460 *** -0.3222 ***
𝜑1(𝑎𝑖𝑗𝑙) Product age 0.0122 * 0.0134 *** 0.0106 ***
𝜑2(𝑎𝑖𝑗𝑙2 ) Squared product age -0.0002 * -0.0003 *** -0.0003 **
Competition 𝛾1(𝐿𝑙𝑚) Consumer loyalty -6.7689 *** -0.3172 *** 𝛾2(𝑆𝑙𝑚) Consumer search 6.4832 ** 0.3068 ** Cannibalization 𝛽1(𝐿𝑙𝑚) Consumer loyalty -13.5006 *** -1.4939 *** 𝛽2(𝐼𝑙𝑚) Consumer income -9.8615 *** -1.0730 ***
𝛽3(𝐿𝑙𝑚∗ 𝐼𝑙𝑚) Consumer loyalty * income 2.2664 *** 0.2559 ***
𝛽4(𝐿𝑙𝑚∗ 𝐼𝑙𝑚2 ) Consumer loyalty * squared income 0.0545 *** 0.0048 ***
𝛽5(𝑆𝑙𝑚) Consumer search 14.3093 *** 1.5700 ***
𝛽6(𝑀𝑗𝑙𝑚) Percent of models offered by firm j 1.3033 *** 0.1384 ***
Observations 2,851 2,851 2,851
Note: * p-value<.05; ** p-value<.01; *** p-value<.001. Standard errors are in parentheses. Due to space limitation, we have omitted all the coefficients for dummy variables included in Equations (10) - (14), including firm dummies, interactions of form factor and channel dummies, CPU category dummies, and yearly dummies.
To estimate the difference in cannibalization effect, we first substitute the estimated form factor, channel, and brand dummies into Equation (11) and calculate the cannibalization effects
σjlm for all the form factor-channel-vendor combinations (listed in Table 4). Then we report the
difference in average σjlm between online and offline channels. Given that we use nonlinear
functions of the estimators to calculate the difference, the p-value is computed based on a Wald test statistic. To estimate the difference in competition effect, we first substitute the estimated
24
form factor and channel dummies into Equation (12) and calculate the four ρlm, one for each form
factor and channel combination (listed in Table 5). We then report the difference in average ρlm
between online and offline channels. Again this difference involves nonlinear functions of the estimators. We compute the p-value based on a Wald test statistic. We find that the online market has both higher cannibalization (difference = 0.0722, p-value < 0.001) and higher competition (difference = 0.0460, p-value < 0.001) than the offline market. This answers our first research question by showing that both cannibalization and competition differ across online and offline markets.
Table 4: Cannibalization Parameter for Each Sub-Nest from Model M1
Vendor Form factor Online Offline Online - Offline
Acer Desktop 0.9249 (0.0165) 0.7583 (0.0252) *** Acer Notebook 0.7574 (0.0047) 0.6292 (0.0337) * Apple Desktop 0.9223 (0.0047) 0.7214 (0.0222) Apple Notebook 0.7459 (0.0594) 0.6629 (0.0718) Dell Desktop 0.9740 (0.0025) 0.8659 (0.0065) *** Dell Notebook 0.8946 (0.0200) 0.6924 (0.0259) Gateway Desktop 0.9653 (0.0140) 0.8621 (0.0200) *** Gateway Notebook 0.8750 (0.0337) 0.8016 (0.0355) * HP Desktop 0.9496 (0.0086) 0.8030 (0.0137) *** HP Notebook 0.8225 (0.0216) 0.7537 (0.0294) * Lenovo Desktop 0.9767 (0.0044) 0.9034 (0.0077) **** Lenovo Notebook 0.8997 (0.0200) 0.8523 (0.0205) * Sony Desktop 0.5055 (0.0107) 0.1878 (0.0111) *** Sony Notebook 0.3038 (0.0807) 0.2367 (0.0069) Toshiba Notebook 0.9175 (0.0395) 0.8726 (0.0427) *
25
Note: * p-value<.05; ** p-value<.01; *** p-value<.001.
Table 5: Competition Parameter for Each Nest from Model M1
Online Offline Online - Offline
Desktop 0.9765 (0.0223) 0.9444 (0.0167) * Notebook 0.9587 (0.0131) 0.8882 (0.0222) ***
Note: * p-value<.05; ** p-value<.01; *** p-value<.001.
Model M2 corresponds to the model specified in Equation (10) with 𝜎𝑗𝑙𝑚 and 𝜌𝑙𝑚
replaced with Equations (13) and (14) respectively. The “Competition” section lists the variables included in Equation (14) that affect inter-brand competition, while the “Cannibalization” section lists the variables included in Equation (13) that affect intra-brand cannibalization. Using this model, we find a similar result to that of Model M1 in that the online market has both higher cannibalization (difference = 0.1428, p-value < 0.001) and higher competition (difference = 0.0557, p-value <0.001) than the offline market.
2.6.2
Consumer Preferences
To answer our second research question, our result of Model M2 shows that both brand loyalty and consumer search play important roles in influencing cannibalization and competition. In terms of cannibalization, the coefficient of brand loyalty is negative and significant (𝛽1 = -13.5006, p-value <0.01), whereas the coefficient for the interaction of brand loyalty and income is positive and significant in both linear (𝛽3= 2.2664, p-value< 0.01) and quadratic terms (𝛽4 = 0.0545, p-value< 0.01). These estimates suggest that brand loyalty has a negative impact on cannibalization if the average income in the market is low, and the impact is positive if the average income is high. This is consistent with our earlier argument that brand loyalties of high-end and low-high-end consumers can have opposite effects on intra-brand cannibalization. It is
26
therefore important to consider consumer composition before drawing conclusions on the effect of brand loyalty on intra-brand cannibalization.
The coefficient for brand loyalty on inter-brand competition, regardless of income, is negative and significant (𝛾1 = -6.7689, p-value<0.01), suggesting that stronger consumer loyalty can mitigate the inter-brand competition. It is consistent with the finding in prior literature that when consumers are more loyal, they become less price-sensitive and less likely to switch between brands (Empen et al. 2011; Krishnamurth and Raj 1991; Raju et al. 1990; Sambandam and Lord 1995).
The coefficient for consumer search is positive and significant for both competition (𝛾2 = 6.4832, p-value<0.05) and cannibalization (𝛽5 = 14.3093, p-value < 0.01). We compare the marginal effects of search on competition and cannibalization and further find that search consistently has a higher impact on cannibalization than on competition in both online and offline markets (p-value<0.01). Accordingly, consumer preference for more prior-purchase search intensifies both inter-brand competition and intra-brand cannibalization. Moreover, more active search by consumers before purchase not only facilitates the comparison of products across firms but also amplifies the substitutability of products within a firm to a larger extent. In addition to the consumer preference variables, the coefficient for 𝑀𝑗𝑙𝑚 is positive and significant (𝛽6 = 1.3033, p-value<0.01), confirming that a relatively long product line tends to induce higher product cannibalization.
As a robustness test, Model M3 shows the results from a 2SLS estimation with linear link functions instead of the logistic transformations in Equations (13) and (14). Our main results hold qualitatively.
27
Now that we have shown that consumer preference variables have significant impacts on both cannibalization and competition, an additional question is: how much variation in
cannibalization and competition between different markets can be explained by the differences in consumer preference variables? Using estimated parameters from Model M2, we calculate the differences in competition and cannibalization effects between online and offline channels as shown in the first row of Table 6. To see the impact of consumer loyalty and search variables, in the second row of Table 6, we assume that there was no difference in loyalty and search
variables between online and offline channels (by replacing their values with the overall average in the market), in which case the differences in competition and cannibalization effects between the two channels come from the channel and form factor dummies only. We can see that our loyalty and search variables explain 53.32% of the difference in competition effect and 48.25% of the difference in cannibalization effect. This finding indicates that our consumer preference variables play a critical role in explaining the cross-market differences in competition and cannibalization.
Table 6: Explaining Power of Loyalty and Search Variables
Online - Offline
Competition Cannibalization
Loyalty and search variables set at observed values 0.0557 *** 0.1428 ***
Loyalty and search variables set at overall average values 0.0260 ** 0.0739 *** Percentage explained by variations in loyalty and search
variables 53.32% 48.25%
Note: ** p-value<.01; *** p-value<.001.
We also compare the consumer preference variables between online and offline channels to draw additional implications on their impacts on the differences in cannibalization and
competition between the two channels. We find that consumers who buy online tend to search significantly more than consumers who buy offline. We do not find a significant difference in consumer brand loyalty across channels, but the interaction between loyalty and income is higher
28
in the online channel due to higher income. This suggests that while the overall higher
competition in the online channel is primarily driven by the difference in consumer search, the higher cannibalization in the online channel is driven both by the difference in consumer search and by the differential effect of brand loyalty of consumers with different preferences for quality. Even though brand loyalty is not statistically different across channels, the higher average
income of the consumers purchasing PCs online suggests a higher consumer preference for quality, ultimately leading to a higher cannibalization in the online channel.
2.6.3
Discussion
Our results indicate that online markets exhibit stronger competition and cannibalization than offline markets. While prior work in information systems has primarily focused on how inter-firm competition differs between online and offline markets, little attention has been paid to the difference in intra-brand cannibalization between online and offline markets. However, both cannibalization and competition are important to firms’ online strategies. Our main result that cannibalization is higher online than offline is consistent with the observation from our data that as a firm increases its number of model offerings, its total sales increase at a lower rate in the online market than in the offline market, as shown in Figure 3.
29
Figure 3: Model Free Evidence of Higher Cannibalization Online
There are at least two reasons why cannibalization can be higher online than offline. First, the online channel has made it a lot easier to search for information. While previous studies have examined the impact of lowered search cost on inter-firm competition, we argue that the lowered search cost, not only affects competition, but also affects cannibalization. More importantly, our result suggests that search in fact can have a higher impact on cannibalization than on competition. The Internet has provided an unprecedented scale of information covering almost every aspect of the product and has also made it very easily accessible. This without doubt intensifies competition between products in the online market, not only across brands but also within the same brand. Because of the amount of details available for the products, it also possibly makes it easier for consumers to distinguish products from different firms and recognize the commonalities between products from the same firm. As a result, while both intra-brand
30
cannibalization and inter-brand competition are higher online, cannibalization could be affected more than competition.
Second, the impact of consumer loyalty on cannibalization can also be different online. Consumer loyalty is not necessarily lower online. The online channel offers features like personalized recommendations and one-click purchasing etc. that can increase loyalty for time conscious consumers. In fact, in our data, we do not find a significant difference in consumer brand loyalty across channels, but the interaction between loyalty and income is higher in the online channel due to the higher income. The people who are more prone to online purchase in the PC market are likely the ones who are more familiar with technology. These people, compared to the general population, can be relatively more knowledgeable and quality conscious. Because brand loyalty of high end consumers has a positive impact on
cannibalization, the characteristics of the online population in the PC market make it more likely for us to observe a higher cannibalization online.
Recognizing the importance of both cannibalization and competition for firm’s online product strategies, our results also show significant economic impact. For example, we can compare the percentage increase in a firm’s total sales after introduction of new products between the online and offline channels through simulations. When we run the simulations, we assume that all the firms are symmetric and all the products are symmetric. We use the mean value for each of our predictor variables to simulate what would happen if one firm increased its number of products. Figure 4a shows the percentage increase in sales as a firm increases its number of products from 16 (which is the average number of computer models per brand in our data) to the number indicated on the X axis. Figure 4b shows the incremental percentage increase in sales if a firm increases its number of products by one assuming its original number of
31
computer models is the number indicated on the X axis. Both plots show a significant higher increase in sales in the offline channel than in the online channel.
(a) What if the focal firm increases the number of products from 16 to a specific number?
(b) What if the focal firm increases the number of products by 1 from a specific number? Figure 4：Comparison between Online and Offline Channels: Percentage Increase in Sales as the
Number of Products Increases
1 1 .0 5 1 .1 1 .1 5 1 .2 1 .2 5 16 18 20 22 24 26 Number of models Offline Online 1 1 .0 2 1 .0 4 1 .0 6 1 .0 8 1 .1 10 15 20 25 30 Number of models Offline Online
32
2.7
Robustness Checks
In this section we explore alternative specifications to check the robustness of our findings. First, intra-brand cannibalization can be affected by the number of computer models. Therefore, instead of using the ratio of the number of models offered by firm j to the total number of models in nest lm, in column M4 of Table 7, we directly use the number of models offered by firm j in nest lm (𝑁𝑗𝑙𝑚) to control for the impact of product variety on cannibalization and use the total number of models in nest lm (𝑁𝑙𝑚) to control for the impact of product variety on competition. Second, as an additional approach to address the concern that search can be endogenous since consumers are likely to search more when there are more products, in columns M5 and M6 of Table 7, we instrument the search variable in our main model using the total number of models in market lm (𝑁𝑙𝑚) and meanwhile we control for the impact of product variety in Equation (13) using either the ratio variable 𝑀𝑗𝑙𝑚 or the number of models 𝑁𝑗𝑙𝑚. Third, to explicitly control for cross-channel competition within the same form factor, in column M7 of Table 7, we modify our two-level nest structure into a three-level nest structure: Level 1 (Form Factor), Level 2 (Channel), Level 3 (Vendor) and computer models at the bottom. As shown in Table 7, our main results hold qualitatively in all the robustness checks.
Table 7: Robustness Checks
M4 M5 M6 M7 Description Number of Models in Cannibalization and Competition IV for Search + Ratio of Models in Cannibalization IV for Search + Number of Models in Cannibalization 3-Layer Model Utility function 𝛼𝑝(𝑝𝑖𝑗𝑙𝑚) Price -0.5282 *** -0.4268 *** -0.5386 *** -0.2687 *** 𝜑1(𝑎𝑖𝑗𝑙) Product age 0.0119 *** 0.0133 *** 0.0085 ** 0.0043 *
𝜑2(𝑎𝑖𝑗𝑙2 ) Squared product age -0.0002 -0.0001 -0.0001 -0.0001
Competition
33 𝛾2(𝑆𝑙𝑚) Consumer search 10.5942 *** 6.0018 ** 7.6573 *** 6.4238 * 𝛾3(𝑁𝑙𝑚) Number of models 0.0091 *** Cannibalization 𝛽1(𝐿𝑙𝑚) Consumer loyalty -13.5496 *** -13.3902 *** -14.1748 *** -13.0430 *** 𝛽2(𝐼𝑙𝑚) Consumer income -10.2586 *** -9.9123 *** -10.5683 *** -10.7237 *** 𝛽3(𝐿𝑙𝑚∗ 𝐼𝑙𝑚) Consumer loyalty * income 2.6050 *** 2.3405 *** 2.5216 *** 2.4542 *** 𝛽4(𝐿𝑙𝑚∗ 𝐼𝑙𝑚2 ) Consumer loyalty * squared income -0.0373 *** 0.0483 *** -0.0535 *** 0.0582 *** 𝛽5(𝑆𝑙𝑚) Consumer search 13.4020 *** 15.2805 *** 15.1942 *** 14.4817 ***
𝛽6(𝑀𝑗𝑙𝑚) Percent of models offered by firm j 1.1467 *** 1.0943 ***
𝛽7(𝑁𝑗𝑙𝑚) Number of models offered by firm j 0.0079 *** 0.0102 ***
Observations 2,851 2,851 2,851
Note: * p-value<.05; ** p-value<.01; *** p-value<.001.
2.8
Conclusions
In this study, we develop a unified GEV framework to simultaneously measure both intra-brand cannibalization and inter-brand competition in online and offline markets as well as the impacts of sear
